Most AI deployments are applied to the surface of a process: the same work, done faster. The ones that compound tend to do something different first. They identify what the existing process is actually built around — the true constraint, not the symptom — and attack that instead.
Four cases from early 2026 where that distinction mattered.
UnitedHealth / OptumRx - Healthcare - Prior authorizations in 30 seconds instead of 8 hours
Prior authorization is healthcare’s most despised bottleneck. The standard framing — shared by insurers, hospitals, and analysts alike — is that approval takes too long because review is slow. Speed up the review, solve the problem.
OptumRx disagreed. The PreCheck AI system processes prior authorizations at intake, before a human reviewer ever sees the request. The previous average was over eight hours; the new average is under 30 seconds. But the number that matters more is this: denials from missing information dropped 68%, and appeals dropped 88%.
The constraint wasn’t review speed. It was missing information at submission. Most prior auth requests were arriving incomplete — wrong codes, absent clinical documentation, missing fields — and the existing process treated that as the reviewer’s problem to resolve. PreCheck fixes the submission before it goes anywhere. Humans only see the cases the model flags as genuinely complex. What looked like a throughput problem was a data quality problem at the front door.
Morgan Stanley - Financial Services - 9 million lines of legacy code that no commercial AI tool could read
Morgan Stanley’s codebase is mostly COBOL and proprietary banking languages that predate every commercial AI coding assistant. When the bank wanted to modernize it, the obvious framing was a talent problem: COBOL expertise is disappearing, and there are not enough developers who can read the old code, let alone rewrite it.
They built DevGen.AI internally because commercial tools were useless against this codebase — but the more important insight was that the constraint wasn’t talent. It was translation. COBOL and proprietary banking languages are alien to modern developers not because development is hard but because the code is unreadable. DevGen.AI converts legacy code into plain English specifications, which any developer can then rewrite in a modern language. Nine million lines processed in 2025.
Solving the translation problem dissolved the talent problem without replacing a single developer. The bottleneck wasn’t the scarcity of COBOL expertise — it was the language barrier that made the existing code inaccessible.
FedEx - Supply Chain - $10M/year saved by predicting equipment failures before they happen
Predictive maintenance is a well-understood AI application. Vendor tools exist. Most large companies evaluating it start with a procurement conversation. FedEx didn’t.
MOBIUS, FedEx’s predictive maintenance platform, reads sensor data from sortation equipment across its surface operations facilities and generates proactive maintenance work orders before failures occur: 17,000 hours of downtime prevented, $10 million saved annually across 41 facilities. They built it internally because the constraint wasn’t willingness to invest in AI — it was data. Commercial predictive maintenance tools are trained on generic industrial equipment. FedEx’s sortation machinery is proprietary, and models trained on other companies’ equipment perform materially worse against it. No vendor had FedEx’s failure history; no vendor could build what FedEx needed.
The constraint wasn’t capability. It was that the only training data that mattered was locked inside FedEx’s own operations. Once that was clear, building internally wasn’t ideological — it was the only option that worked.
JPMorgan Chase - Financial Services - The bank tracking which engineers use AI, and which don’t
JPMorgan’s 65,000-person Global Technology division is now operating under explicit AI adoption objectives. Engineers are expected to show measurable improvement in code quality, speed, and productivity through regular use of approved AI coding tools. One internal dashboard tracks GitHub Copilot usage and classifies every developer as a “light,” “heavy,” or “non” user — by name.
This looks like an accountability program. The underlying insight is a constraint identification: most companies deploying AI have no idea whether the deployment is actually being used. Adoption is invisible. Rollouts happen, licenses get bought, and the productivity gains that were supposed to materialize either do or don’t — and nobody can say with precision why.
JPMorgan named the invisible thing and measured it. The constraint on AI ROI at the deployment scale they’re operating isn’t the technology — it’s knowing whether the technology is being engaged with at all. The dashboard answers that question. Most companies are still guessing.
Throwback: Capital One, 1994 — Credit Was Never a Judgment Problem
In 1994, Rich Fairbank and Nigel Morris convinced Signet Bank to spin out a credit card company built around a single contrarian idea: that credit approval — universally treated as a judgment call by experienced underwriters — was actually a data problem. If you had enough history on how borrowers behaved, a model would outperform the rules committee.
Capital One spent the next two decades being right. By 2015, ML models were running real-time credit decisions for approximately 65 million customers — not assisting underwriters, running the decisions — built on a data asset compounded over 20 years of lending history. Fairbank described Capital One in shareholder letters as a technology company that happened to have a banking charter.
The transformation was invisible while it happened. No press release announced the replacement of the underwriting rules committee with gradient boosting. Capital One just ran the models, compounded the data advantage, and let the credit performance speak. The “judgment” that everyone said couldn’t be systematized was pattern recognition on historical data all along. The industry spent thirty years arguing about whether credit decisions could be automated. Capital One spent those same thirty years automating them.
OptumRx fixed the missing information, not the review. Morgan Stanley dissolved the language barrier, not the talent gap. FedEx built what only FedEx’s data could build. JPMorgan measured what everyone else was guessing at. Capital One did all of this in 1994, for a problem the whole industry had convinced itself was fundamentally human. The pattern is the same across six decades: the constraint is almost never the one the existing process was designed to solve.